Depth Estimation Model Using Perceptual and Color Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing depth estimation methods suffer from inaccuracies in low-texture areas due to small color differences between pixels, leading to misjudgment and reduced accuracy of depth information.
Innovation Solution
A depth estimation method utilizing an autoencoder-based model that incorporates a mean square error and cosine similarity calculation to refine pixel value predictions, with a training process that updates parameters using a back-propagation algorithm to improve prediction accuracy and color sensitivity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing depth estimation methods are used, then depth information can be obtained through image reconstruction, but accuracy deteriorates in low-texture areas due to small color differences between pixels
Solution Approach 1:
The patent changes the loss function parameters from traditional L1/L2 loss to a combination of perceptual loss (based on pre-trained network features) and color consistency loss (based on LAB color space). This parameter change enables the model to capture subtle color differences in low-texture areas while maintaining overall depth estimation accuracy, directly resolving the contradiction between measurement precision and reliability in challenging regions
Solution Approach 2:
The patent introduces LAB color space as an intermediary representation to bridge the gap between RGB image data and depth estimation. By converting RGB images to LAB color space and using the a and b channels (which contain color information) for training, the method enhances the model's ability to detect subtle color variations in low-texture areas, thereby improving both accuracy and reliability simultaneously
2Measurement precision
If traditional loss functions are used for training, then training process is simple, but color sensitivity and depth estimation accuracy are insufficient
Solution Approach 1:
The patent segments the loss function into multiple independent components: perceptual loss (using pre-trained network features) and color consistency loss (using LAB color space). This segmentation allows each component to specialize in different aspects - perceptual loss captures structural information while color consistency loss captures color variations - thereby enhancing color sensitivity without creating an insurmountably complex training model
Solution Approach 2:
The patent performs preliminary conversion of RGB images to LAB color space before training, and pre-trains a network to extract perceptual features. These preliminary actions prepare the data and model in advance, making the main training process more efficient despite the enhanced complexity of the loss function, thus achieving high color sensitivity with manageable training complexity
Data Source
AI summary
A method for training a depth estimation model implemented in an electronic device includes obtaining a first image pair from a training data set; inputting the first left image into the depth estimation model, and obtaining a disparity map; adding the first left image and the disparity map, and obtaining a second right image; calculating a mean square error and cosine similarity of pixel values of all corresponding pixels in the first right image and the second right image; calculating mean values of the mean square error and the cosine similarity, and obtaining a first mean value of the mean square error and a second mean value of the cosine similarity; adding the first mean value and the second mean value, and obtaining a loss value of the depth estimation model; and iteratively training the depth estimation model according to the loss value.